AI Models From ANY Angle & SUPIR Upscaling (Low VRAM Method)

hey there in today's video we're going to learn how to get amazing images of an AI Digital model from different angles while ensuring full consistency plus we'll be upscaling our images with the topnotch image restorer and upscaler superp now a quick heads up everything you'll see in this video was created using an RTX 360 with 12 GB of vram from generating images with IP adapter to upscaling them using superer this GPU handled it all I'll explain each step and every node in detail and you'll find all the resources in the description box if you're new here don't forget to like And subscribe now let's get [Music] [Applause] [Music] started first things first the checkpoint model we're using today is called level four XEL it's a turbo model meaning it generates images super fast and it's a versatile hybrid model capable of creating realistic shots anime CGI and 3D digital art much like dream shaper XL you can find level four XL on civid AI and the recommend settings are between 8 and 16 steps with a CFG scale between 1 and three make sure to use the DMP Plus+ SDS sampler and set the Schuler to caras now now let's fire up comfy UI and load a default workflow our challenge is to generate a character's face from multiple angles keeping the facial features consistent first we'll select our checkpoint model for the positive prompt to vary the facial features you can use descriptions such as Spanish or German nationalities you can also specify age to get a young or older character for the negative prompt you can filter out elements you don't want want like jewelry or anything else choose any supported sdxl dimensions for your image resolution I'm using 832 X 1216 pixels we've already checked the recommended settings for the K sampler so we'll stick with that if you're using a different checkpoint model make sure to check their recommended settings on their civid AI page to get the best results all right let's get started and make some compelling character face shots now let's get our hands dirty and create a single image featuring four character faces from different angles to do this we'll use this reference image which you can find Linked In the description below this image contains four 3D grid of heads from various perspectives next we'll utilize the depth anything tool and connect it to an apply control net Advanced node with the Zoe depth model for sdxl [Music] running this workflow will generate our characters faces in the specified directions provided by the depth control [Music] net feel free to tweak the prompts to adjust features or fix any issues in the generated images once you're happy with the result it's time to upscale the image for better quality using ultimate SD upscale we'll test four different upscaler models to compare the final results and choose the best one keep the ultimate SD upscale node settings as recommended by the model you're using duplicate these notes three times by holding control shift and pressing V to download upscaling models head over to openm modelb doino you can search for any upscaler and download it for free the site also provides live examples of each model's capabilities and additional [Music] information one that downloaded place the model in the upscale models folder within the models directory inside comfy UI after the upscaling is complete compare the images to see which model delivered the best results [Music] while the differences may be minor I find that the ultra sharp model produces the best skin tones remember take your time with this step the more you refine and experiment the better your consistent character will be for future projects all right now let's move on to separating the four Images since we upscaled our image by four each individual image will be 832 by 1216 pixels to do this we use just one node the image crop node first to get the initial image in the top left of the grid set the dimensions to 832 pixels in width and 1216 pixels in height choose the position as top left for the other three images we'll duplicate this node three times simply select the node copy it with control+ C then paste it by holding contrl plus shift and pressing V now let's adjust their positions the second image should be top right the third image should be bottom left the fourth image should be bottom right and just like that we've cropped the upscaled image into four individual images each showing our character from different angles first remove the ups scalers and control n nodes keeping only the basic elements for our positive prompt let's go with something like realistic photograph and different face angles add a happy emotion like smiling we'll set the background to a simple City Park you can also specify clothing and lighting if you'd like next we need a batch of four Images load the IP adapter and the unified loader connect these two noes and select the plus face portrait model for the I adapter remember to lower the weight a bit now to load our four face images use the load image batch from dire node from the Inspire pack this node will load all the images from a specified folder so create a new folder for your face images copy its path address and paste it in the node connect the model and IP adapter to the case sampler and let's generate the [Music] images check this out we have four images with consistent faces clothing and backgrounds the characters are smiling too however only one image shows a different face angle the rest look kind of similar with a weight value of 0.6 the characters don't look exactly like our reference images by increasing the weight to 0.8 the digital model's face will look more like our reference but this may affect the emotion and keep the face angle the [Music] same to fix this we can use another phas processor like face ID along with plus face but note that face ID works with inside face and isn't allowed for commercial use that's why in my last AI Digital model video I only use the plus face IP adapter if you haven't watched that video make sure to check it out after this one first load the face ID unified loader and the face ID nodes connect them directly to our model and then chain them with the IP adapter and plus face set the weight to 0.75 for face ID and 0.6 for plus face make sure you select face ID plus face version 2 a quick note this node won't run without inside phas I'll include step-by-step instructions on how to install it on Windows along with all necessary resources and links in the description below now when generating your batch images you'll start noticing the improved impact of face ID version 2 your character's face will have more life with shots taken from different angles the similarity to your reference images will also be better by tweaking the face ID weight and allowing some Freedom where the IP adapter ends you can balance your character's face features using the same parameters and reference images ensures you get the same face every time you generate a new image [Music] now let's dive into my favorite method for upscaling images superpier this tool is fantastic for restoring images using text prompts and does an amazing job but keep in mind this method requires a lot of GPU power the good news with the right setup and a models you can use super smoothly with just 12 GB of vram or more first load the workflow provided in the description box make sure to install any missing nodes from The Comfy UI manager you'll need the impact pack ultimate SD upscale RG3's nodes comfy UI Essentials KJ nodes and of course superpier for comfy UI just a heads up using a portable version of comfy UI might cause nodes to conflict I fixed this by doing a fresh install of comfy UI using Pinocchio since then I haven't had any issues and can load any workflow after setting up the nodes from the manager I'm not sure why ultimate SD upscale or RG3's nodes don't load well on the portable version of comfy UI but Pinocchio has been my go-to solution for ensuring all nodes install correctly next download the superp pruned checkpoint models from the hugging face link below and place them in your checkpoint folder you'll also need an sdxl model for the best results I recommend using Juggernaut XL version 9 lightning and turbo models work too but they can compromise image quality after restarting comfy UI and loading the sup workflow let's break down what we see the group nodes labeled 2K by sup will upscale your image to approximately twice its original size on the right there's the ultimate SD upscale which will take that already upscaled image and boost it up to 8K resolution if the ultimate SD step is too demanding you can skip it and stick with just the super 2K group [Music] node you can adjust your image height here setting this value to 4,000 will give you a 4K image but be aware this will slow down processing time unless you have over 24 GB of VR in our example we'll use both the 2K super and ultimate SD to see the difference for the right setup start by loading Juggernaut XL version 9 in the super model load node select Super's trained model v0 qfb 16 if you have a lower vram choose the bf16 encoder instead of Auto for both super first stage and super encode set the super sampler to 40 steps you can also choose a different upscaler model for the ultimate SD upscale before generating your image for the positive prompts you can guide the model with your image description but the default prompts work just fine now let's upscale the image and see how long it takes to generate a 2K and 8K image [Music] using a 12 GB card the entire process took 3 minutes and 20 seconds as you can see the image quality and texture have significantly improved the facial features of our characters are restored without losing any likeness this is the true power of superpier and that's it for today's tutorial if you encounter any errors feel free to send a screenshot to the email in the description below and I'll try to respond as soon as possible don't forget to like share and subscribe if you haven't yet see you in the next video [Music]

Channel: Aiconomist

Views: 12,421

Keywords: comfyui, stable diffusion, ai art, workflow, automatic 1111, forge, midjourney, ai

Id: AyjdjF0YP_s

Length: 14min 23sec (863 seconds)

Published: Sun Jun 02 2024